In an era where journals are under mounting pressure to implement double-blind peer review while handling rapidly increasing submission volumes, editorial offices still depend largely on manual redaction or coarse “document inspector” tools to remove identifying details from manuscripts and reviewer reports. These practices are labor intensive, difficult to standardize across editors, and often act as blunt instruments that disrupt the review process by removing useful layout metadata together with sensitive information. They also provide limited protection against implicit semantic leakage, To mitigate these risks and reconcile the tension between robust privacy protection and editorial efficiency, this study proposes an intelligent bidirectional privacy anonymization strategy that integrates rule-based algorithms with large language models (LLMs) and implements it as a scalable browser/server application aligned with editorial workflows. Grounded in an analysis of typical submission materials, the system formalizes three design dimensions: supported file formats, sensitive information categories and high-risk document locations. It supports mainstream word-processing formats, targets core identifiers for authors and reviewers, and concentrates on predefined high-risk locations. On this foundation, we construct a two-layer hybrid engine. A rule-based layer, implemented against the Office Open XML schema, uses regular expressions and structural cues to deterministically locate and neutralize well-structured fields such as author lists, affiliations and email addresses while explicitly protecting in-text citations and reference lists as spans that must not be altered. An LLM-based layer is then invoked through structured prompts that encode editorial heuristics and send only minimal, context-tagged text segments to the model. This layer identifies and masks residual identity cues that escape rule-based detection—the "long tail" of semantic leakage. For PDF files, whose internal structure is less amenable to safe in-place editing, the system adopts a non-destructive “sensitive-information warning” mode in which extracted text is screened and suspected identifiers are flagged for manual verification rather than being automatically rewritten. The hybrid approach is extended symmetrically to reviewer reports. For DOC/DOCX files, the system parses comment and revision nodes and replaces user names and contact details with neutral labels such as "journal editor" or "reviewer A" while preserving the original review content; for PDF reports, suspected identity fields are similarly highlighted for anonymization or human confirmation. The anonymization engine is exposed through a web interface and standardized application programming interfaces, enabling on-demand use by editors and integration with editorial management systems at key workflow stages. An internal evaluation of real manuscripts and reviewer reports by experienced editors indicates that the rule-plus-LLM strategy more reliably removes explicit identifiers and reduces implicit identity cues in high-risk locations than manual or rule-only approaches, without altering in-text citations or reference lists, and substantially shortens the preparation time for double-blind review. Comparison of manual, rule-only, LLM-only and hybrid schemes suggests that the proposed engine achieves a favourable balance of precision, coverage, consistency and operational cost. Overall, this study demonstrates the feasibility of deploying a rules-plus-LLM hybrid engine for intelligent bidirectional anonymization in journal peer review and offers a practical, scalable pathway for journals seeking to strengthen privacy protection and editorial efficiency, while building a fairer and more trustworthy peer-review ecosystem.